Latent Conjunctive Bayesian Network: Unify Attribute Hierarchy and Bayesian Network for Cognitive Diagnosis

Abstract

Cognitive diagnostic assessment aims to measure specific knowledge structures in students. To model data arising from such assessments, cognitive diagnostic models with discrete latent variables have gained popularity in educational and behavioral sciences. In a learning context, the latent variables often denote sequentially acquired skill attributes, which is often modeled by the so-called attribute hierarchy method. One drawback of the traditional attribute hierarchy method is that its parameter complexity varies substantially with the hierarchy's graph structure, lacking statistical parsimony. Additionally, arrows among the attributes do not carry an interpretation of statistical dependence. Motivated by these, we propose a new family of latent conjunctive Bayesian networks (LCBNs), which rigorously unify the attribute hierarchy method for sequential skill mastery and the Bayesian network model in statistical machine learning. In an LCBN, the latent graph not only retains the hard constraints on skill prerequisites as an attribute hierarchy, but also encodes nice conditional independence interpretation as a Bayesian network. LCBNs are identifiable, interpretable, and parsimonious statistical tools to diagnose students' cognitive abilities from assessment data. We propose an efficient two-step EM algorithm for structure learning and parameter estimation in LCBNs. Application of our method to an international educational assessment dataset gives interpretable findings of cognitive diagnosis.